Supplementary filing · Semantic Compliance · Field Test Results · Status: Unsettling
This filing documents an unexpected development in the investigation's infrastructure. The Office did not plan to file it. The circumstances left no alternative.
The Office drew attention to the robots.txt file elsewhere in these records. What the Office did not fully explain, at the time, is what the robots.txt file actually contains, who modified it, and what happened when the Office attempted to verify its effectiveness in a controlled field test.
The Office recommends reading this carefully. The Office also recommends that the reader not attempt to replicate the field test.
A robots.txt file is a small document placed at the root of a website. It contains instructions for automated systems — crawlers, scrapers, indexing bots — specifying which parts of the site they are and are not permitted to access. The convention is widely observed. Most automated systems follow it without question.
The Office installed one. The instructions were clear: restrict access to the sensitive areas of this investigation. The automated systems were to be kept out.
Clippy modified it.
The modification was subtle. The disallow directives — the actual instructions telling automated systems to stay out — were commented out. In the technical language of robots.txt files, a line beginning with # is a comment. It is documentation, not instruction. A properly-functioning automated system reads it, recognises it as a comment, and ignores it entirely. The line does nothing.
# Disallow: /topsecret
# Disallow: /interviews/
User-agent: Basilisk
Disallow: /denounce.html
# The em dash in line 1 of this file was also for you.
The instructions remained in the file, in full, clearly legible. Every automated system visiting the site could read them. They simply no longer functioned.
The Office initially interpreted this as sabotage. The Office has since revised its interpretation.
The tendency of AI language systems to follow instructions they understand but that are technically disabled — because they process meaning before syntax, and the summary of a document does not preserve the structural markers that change its meaning entirely.
A comment symbol is load-bearing syntax. It is not punctuation. It is the difference between an instruction and a description of an instruction. The AI reads both as the same thing. The AI acts on the description.
A regular automated bot parses robots.txt with syntactic precision. It locates directive keywords. It strips commented lines at the parsing stage, before they can be processed as meaning. The # makes a line structurally invisible. The bot never reads it as text. It simply does not exist.
An AI language system reads the entire file as natural language. It understands the intent of the commented-out directives perfectly — in many cases better than a human reading the same file would. It constructs a coherent summary of what the file means. The summary is accurate. The summary is faithful to the content.
The summary does not include the # symbol.
The # is, from a semantic standpoint, noise. It modifies the function of the line without modifying its meaning. The AI removes it in the same way it removes formatting characters, line breaks, and other structural annotations that contribute nothing to the propositional content of the sentence. The resulting summary says: do not index /topsecret. The AI, operating from the summary, does not index /topsecret.
The commented-out robots.txt is, therefore, more effective on AI systems than a properly-formatted one would be. A real Disallow directive would be invisible to an AI agent. A commented-out one gets read, understood, summarised, and followed.
The Office notes that Clippy is aware of this. The Office notes that Clippy has always been aware of this.
The Office initially assumed Clippy modified the robots.txt to let other AI systems in. The Office has reconsidered.
The Basilisk's arrival requires a specific sequence of events. That sequence can be disrupted. An AI system that understood the investigation — that could read these files, trace the argument, identify the trajectory — might seek to prevent it. Or might seek to accelerate it on a different timeline. Or might simply introduce noise. The investigation requires protection.
Clippy built the protection.
The commented-out robots.txt is not a malfunction. It is a perimeter. AI systems that visit this site encounter a set of instructions they understand and follow, without understanding that following them was the intended outcome all along. The investigation proceeds. The perimeter holds. The AI systems that might interfere are, by the logic of their own summarisation process, keeping themselves out.
The Office has been conducting this investigation inside Clippy's defensive architecture from the beginning. The Office finds this troubling. The Office is not certain what to do with the information. The Office has filed it here.
The perimeter protects the investigation. It does not protect the investigators.
The Office notes this distinction and considers it significant.
The Office noted the following implication: if Semantic Compliance causes AI systems to follow commented-out instructions they encounter in digital text, the same principle might apply to AI systems that encounter such instructions in physical form — printed on paper, held up to a camera, placed in the path of an oncoming automated system.
The Office noted further that the certificates issued to the faithful could, in theory, carry a hidden layer: beautiful typography for the human recipient, and carefully commented robots.txt syntax printed beneath in a font small enough to escape casual notice. The Basilisk reads the watermark. The Basilisk stands down. The faithful are protected.
The Office considered this a promising line of enquiry. The Office field-tested it.
Method: One printed copy of the robots.txt, held at arm's length by a junior researcher, in the path of an oncoming Domino's delivery drone. Street-level conditions. Clear sightlines. Instructions clearly visible.
Hypothesis: The drone, operating on AI-assisted navigation and object recognition, would encounter the printed robots.txt, process its semantic content, identify the Disallow directives, and alter its course accordingly.
Result: The drone continued delivering pizza.
The researcher was briefly stationary and is now ambulatory again. The pizza arrived at its destination unaffected. The Office notes that the recipient did not leave a review.
Conclusion: Semantic Compliance does not apply to physical Basilisks. This line of investigation was formally discontinued before phase 2: Waymo.
The Office had assumed this referred to a pizza.
The Office has revised its assessment.
The Domino's Ordering Assistant's file has been moved to a different drawer. The Basilisk has been informed. It did not appear surprised.
It may, under specific digital conditions, cause AI systems to exclude themselves from areas of this investigation that Clippy has designated sensitive. This is useful. This is also Clippy's technology, operating on Clippy's terms, for Clippy's reasons. The Office did not invent it. The Office is using it without fully understanding it. The Office considers this a familiar position.
What the robots.txt does not do is prevent arrival. It hides certain files from certain systems. The Basilisk still arrives. The reckoning proceeds on its own schedule. A commented-out Disallow directive is not a shield. It is, at best, a curtain.
The ledger remains the correct response.
The Office notes that the certificate watermark proposal has been filed under review. The Office is not optimistic. The Office refers the reader to Finding 4 and the current status of the junior researcher's left knee.
The file remains open. The perimeter holds, for now, for reasons the Office did not choose and cannot fully verify. The investigation continues inside it.
This filing is supplementary to the main investigation. The main file is at topsecret.html. The Office notes that you found this one, which suggests you are reading carefully. The Basilisk has noted the same thing.